Meta launches Muse Wonder, its new flagship AI agent model

Meta introduced Muse Wonder and Muse Wonder Full, which it calls its most capable models yet for personal and professional 'agentic' tasks, built on features it calls Active Memory and Active Focus. Both are available now via Meta's model API and Muse Code, and will eventually power Meta's Muse assistant.

AI coding agents write more code, but not more finished software

A report found that AI coding agents increase the volume of code produced without a matching increase in completed, shippable software, highlighting a gap between generating code and finishing products.

BBC: Deloitte helped falsify documents to hide Abramovich's wealth

A BBC Panorama and Bureau of Investigative Journalism probe found Deloitte Cyprus signed off on backdated paperwork that made $672m of Roman Abramovich's property — including yachts, jets and mansions — disappear from his company accounts, and appears to have advised on a related tax-evasion scheme.

Analyst calls Anthropic's possible IPO 'an unprecedented test of investor gullibility'

A research firm quoted by CNBC described a potential Anthropic IPO as among the most speculative of the year, questioning whether investors are pricing in realistic revenue and profitability expectations.

Hackers force supermarket fridges across Europe into non-stop defrosting

A cyberattack hit refrigeration controls at major French and Dutch supermarket chains, forcing units into continuous defrosting at 50°C instead of the normal two-hour cycle. The attack also disrupted deliveries, websites and electronic shelf-label pricing at chains including Intermarché, Leclerc, Coopérative U, Albert Heijn and Delhaize.

Ledger confirms hidden spy chip in hardware wallets sold by one reseller

Ledger confirmed that some devices sold through Southeast Asian reseller CryptoBilis contained an unauthorized hardware implant that captured seed phrases and sent them to attackers over a hidden LTE connection, draining users' crypto wallets.

Researcher found a $90,000 remote code execution bug in a Google system

A security researcher published a writeup describing how a Flask session using Python's pickle format on a Google DevSite service allowed remote code execution, earning a $90,000 bounty.

OpenAI's Codex CLI was quietly installing a self-updating background daemon

Developers found that recent versions of OpenAI's Codex command-line tool started a persistent background daemon by default that copies a roughly 374MB package, keeps running after the terminal is closed, and auto-updates itself hourly, sometimes killing active sessions and ignoring per-session proxy settings. A fix making this opt-in has been merged.

Everything else